Stefan wrote: > Am Sonntag, 9. August 2009 07:36:54 schrieb Luis Daniel Lucio Quiroz: >> Hi SAs, >> >> Well, after reading this link >> http://spamassassin.apache.org/full/3.2.x/doc/sa-learn.html I'm still >> looking for an easy-way to let my mortal users to train our antispam. I >> was thinking a mailbox such as h...@antispamserver and s...@antispamserver >> to let users to forward their false positivos or their false netgatives. >> In isde each box (ham or spam), of course a procmail with sa-learn input >> will be forwarded. >> >> My doubts are nexts: >> 1. Will forwarded mails be usefull for training, I mean if spam was: From: >> spa...@example.net To: u...@mydomain, when forwarding it will be From: >> mu...@mydomain To: s...@antispamserver. Change of this and forwarding >> (getting rid of headers because mail-clients) wont change learning? > > You have to forward the message as an attachment un unpack it after > receiving. > Have a look at: > https://po2.uni-stuttgart.de/~rusjako/sal-wrapper
Yes, I find this approach works well. It's the simplest way for me to train Bayes, and most users can cope with it, providing they're not using Outlook 2003/XP which can't forward as an attachment. But Thunderbird, Outlook Express, Squirrelmail and Pine all can easily. It's not as simple as a 'This Is Spam' button perhaps, and that's a *good* thing. Requiring a little bit of thought stops people using it as an alternative to the delete key for 'OK, perhaps I did subscribe to this but I don't want it now'. My script is very similar to sal-wrapper, using Postfix check_recipient_access to ensure only authenticated users can send to the reporting address; triggered from procmail; using MIME::Parser to extract (possibly multiple) message/rf822 attachments; feed through sa-learn --ham or spamassassin -r as appropriate and send an acknowledgement back to the user, to remind them to also send spam/non-spam to the corresponding address and correct any mistakes. One thing I notice from sal-wrapper however is that it pipes the header and body to sa-learn without passing a file as parameter. I found that although sa-learn didn't complain, this didn't work at all well, and quite short ham messages were scoring BAYES_99. You can pipe to spamassassin -r just like you can to spamassassin in any other mode, but I think if you pipe to sa-learn, you need to do it as sa-learn --ham - with the '-' as parameter, so it reads the standard input. Alternatively feed it a temporary message file. Or am I misreading something? CK
